• No results found

The association of overweight and obesity with incident heart failure

N/A
N/A
Protected

Academic year: 2020

Share "The association of overweight and obesity with incident heart failure"

Copied!
182
0
0

Loading.... (view fulltext now)

Full text

(1)

THE ASSOCIATION OF OVERWEIGHT AND OBESITY

WITH INCIDENT HEART FAILURE

Laura R. Loehr MD, MS

A dissertation submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the

Department of Epidemiology from the School of Public Health.

Chapel Hill 2008

Approved by,

Advisor: Wayne D. Rosamond PhD, MS Reader: Gerardo Heiss MD, PhD

Reader: Charles Poole ScD

Reader: Patricia P. Chang MD, MPH

(2)

ii @ 2008

(3)

iii

ABSTRACT

Laura R. Loehr, MD MS: The association of overweight and obesity with incident heart failure

(under the direction of Wayne D. Rosamond, PhD MS)

Obesity has been identified as a risk factor for heart failure. The importance of a measure of central adiposity (waist-hip ratio) as compared to BMI has not been extensively studied. The increasing prevalence of both obesity and heart failure (HF) make this association an important topic for prevention.

(4)

iv

With stratification by race and gender, the adjusted hazard ratios for the comparison of the highest category of each anthropometric measure (obese) to the lowest were similar and ≥

1.0. Results from the sensitivity analysis estimated the effect of outcome misclassification was to bias our findings toward the null. Calculation of the generalized impact fraction estimated that a hypothetical 30 % reduction in the prevalence of obesity would reduce incident HF by 6.7 % in the population.

(5)

v

ACKNOWLEDGEMENTS

I would like to express my sincere appreciation to Dr. Wayne D. Rosamond, my

committee chair, for his enduring patience and excellent mentoring over the last few years. This dissertation would not have been possible without his guidance. I would like to thank my committee members for their many contributions: Drs. Gerardo Heiss, Charles Poole, Ann Marie McNeill, and Patricia P. Chang. In addition, thanks to co-authors Drs. Aaron Folsom and Lloyd Chambless for their contributions to manuscript 1. I have been very fortunate to have such a great team of people to work with. I much appreciate the friendship of everyone in the CVD epidemiology group, and especially Phyllis Johnson.

(6)

vi

TABLE OF CONTENTS

LIST OF TABLES ... ix

LIST OF FIGURES ... xii

LIST OF ABBREVIATIONS ... xv

CHAPTERS I. SPECIFIC AIMS ... 1

II. BACKGROUND AND SIGNIFICANCE ... 4

A. Public Health Burden of Heart Failure and Obesity ... 4

B. Risk Factors for Heart Failure... 6

C. Obesity as a Risk Factor for Heart Failure ... 7

D. Studies of Obesity as a Risk Factor for Heart Failure ... 11

E. Validation Studies of ICD codes to Define Heart Failure... 20

F. Methods to Adjust for Bias... 28

G. Impact measures of risk factor-disease associations ... 32

H. Summary and public health significance... 35

III. METHODS ... 38

A. Study population ... 38

B. Exclusion criteria... 39

C. Ascertainment of heart failure events ... 40

D. Incident heart failure event criteria ... 41

(7)

vii

F. Baseline covariate definitions ... 42

G. Data quality ... 43

H. Statistical power analysis ... 44

I. Statistical Analysis... 47

J. Methodological strengths and limitations... 62

IV. RESULTS... 64

A. Manuscript 1: The Association of Overweight and Obesity with Incident Heart Failure: the Atherosclerosis Risk in Communities (ARIC) Study ... 64

1. Introduction... 65

2. Methods... 66

3. Results ... 72

4. Discussion... 75

B. Manuscript 2: The preventable burden of heart failure due to obesity: The Atherosclerosis Risk in Communities (ARIC) Study ... 87

1. Introduction... 88

2. Methods... 90

3. Results ... 97

4. Discussion... 99

V. CONCLUSIONS ... 110

A. Recapitulation of overall specific aims... 110

B. Discussion of results ... 112

C. Degree to which doctoral goals have been met... 115

D. Strengths ... 117

E. Limitations ... 118

(8)

viii

APPENDICES ... 122

A. IRB certification... 123

B. Supplemental Results, Manuscript 1... 125

C. Supplemental Results, Manuscript 2... 153

(9)

ix

LIST OF TABLES

Table 1. Summary table of studies of obesity as a predictor of heart failure ... 19

Table 2. Summary table of validation studies of ICD codes to define heart failure... 27

Table 3. An example of the succinct summarization of results from a semi-automated ... 31

Table 4. Description of Gothenburg score... 40

Table 5. Two group test of equal exponential survival, with exponential dropout, only men included... 45

Table 6. Two group test of equal exponential survival, with exponential dropout, women only ... 45

Table 7. Estimate of power* to assess for multiplicative effect measure modification by the following variables, given probabilities and sample size in the ARIC... 47

Table 8. (MS. 1, Table1) Characteristics at baseline(1987-1989) of those who did or did not develop heart failure, ARIC ... 81

Table 9. (MS. 1, Table 2) Number of heart failure cases, total person-years of follow-up and age-adjusted* incidence rates (IR) for heart failure by category of each anthropometric measure (BMI, waist circumference, and waist-hip ratio), stratified by race and gender, ARIC, 1987-2003... 82

Table 10. (MS. 1, Table 3) Unadjusted and adjusted* hazard ratios (with 95 % CI) for incident heart failure by BMI, waist circumference and waist-hip ratio, stratified by race and gender, ARIC, 1987-2003... 83

Table 11. (MS. 1, Table 4) Numbers of heart failure events, person-years of follow-up, age-adjusted heart failure incidence rates and multivariable age-adjusted hazard ratios of BMI stratified by high or low waist-hip ratio (WHR)Ω, ARIC 1987-2003... 85

Table 12. (MS. 2, Table 1) Characteristics of the population (N = 14,642) at baseline by category of BMI, ARIC, 1987-1989 ... 105

Table 13. (MS. 2, Table 2) Total number of participants, number of heart failure events, follow-up time, proportion in each BMI category, and unadjusted hazard ratios for incident heart failure, stratified by race, gender and age, ARIC 1987-2003 ... 106

(10)

x

(11)

xi

(12)

xii

LIST OF FIGURES

(13)

xiii

(14)

xiv

(15)

xv

LIST OF ABBREVIATIONS

AF Attributable fraction ANOVA Analysis of variance

ARIC Atherosclerosis Risk in Communities BMI Body mass index

BNP B-type natriuretic peptide

BP Blood pressure

CAD Coronary artery disease CHD Coronary heart disease CHS Cardiovascular Health Study

CI Confidence interval

CIE Change in estimate

CLR Confidence limit ratio

CT Computed tomography

CVD Cardiovascular disease DAG Directed acyclic graph

FEV1 Forced expiratory volume in 1 second GIF Generalized impact fraction

HERS Heart and Estrogen/Progestin Replacement Study

HF Heart failure

HDL-C High-density lipoprotein cholesterol HOPE Heart Outcomes Prevention Evaluation

(16)

xvi ICD International Classification of Disease

ICD-9-CM International Classification of Disease, 9th revision, clinical modification ICD-10-CM International Classification of Disease, 10th revision

ICR Interaction contrast ratio

IR Incidence rate

JAMA Journal of the American Medical Association LDL-C Low-density lipoprotein cholesterol

LRT Likelihood ratio test

LVH Left ventricular hypertrophy

MD Maryland

MI Myocardial infarction

MN Minnesota

MS Mississippi

MSE Mean square error

N Number

NIH National Institute of Health

NC North Carolina

NHANES (I) National Health and Nutrition Examination Survey Epidemiologic Follow-up Study

(17)

xvii

RR Relative risk

P Proportion

PHA Proportional hazards assumption

PY Person-years

SD Standard deviation

SE Standard error

T1 First tertile

T2 Second tertile

T3 Third tertile

WC Waist circumference

(18)

CHAPTER I

SPECIFIC AIMS

This study aims to fulfill the following objectives: 1) quantify the association between overweight and obesity as measured by waist/hip ratio, waist circumference, and BMI with incident hospitalized heart failure; 2) compare the magnitude of association for these three anthropometrical measures as to their ability to predict heart failure; 3) assess the probable magnitude and direction of systematic error due to outcome misclassification; 4) determine the effect of outcome misclassification on the association between obesity and heart failure using a semi-automated probabilistic sensitivity analysis; 5) determine the population burden of incident heart failure that could be prevented if there were a hypothetical reduction in the distribution of obesity and overweight.

Specific Aim 1

Assess obesity/overweight as risk factors for the development of incident hospitalized heart failure. This aim will be achieved by the following sub-aims:

(19)

2

1.2: Compare three anthropometrical measures (e.g. BMI, waist circumference, and waist/hip ratio) as to their ability to predict incident hospitalized heart failure.

Specific Aim 2

Assess the probable magnitude and direction of systematic error due to misclassification of the outcome of heart failure for the association of obesity with incident heart failure. This aim will be achieved by the following sub-aims:

2.1: Determine a distribution for sensitivity and specificity as estimated from the literature for the definition of the outcome (incident HF) as defined by hospital discharge and death codes.

2.2: Perform a semi-automated probabilistic sensitivity analysis to estimate the degree of bias due to disease misclassification based on the chosen distribution of sensitivity and specificity. Separate multivariable estimates of the odds ratio and its distribution will be obtained that include systematic error (from disease misclassification), random error and both.

Specific Aim 3

(20)

3

3.1 Determine a range for the feasible reduction in the prevalence of obesity and overweight (as measured by BMI), based on the findings from the literature.

3.2 Given several hypothetical scenarios based on the literature for the reduction in prevalence of obesity and overweight, and the magnitude of association between BMI and incident heart failure (as determined from the previous aims), the generalized impact fraction will be determined, overall and stratified race, gender, and age. This will estimate the potential population-level impact of weight reduction on the incidence of heart failure.

3.3 Determine the population attributable fraction for overweight/obese for the outcome of incident hospitalized heart failure. Compare this more commonly used measure to that of the generalized impact fraction.

Rationale

This study adds to the existing literature on obesity and heart failure in several ways. First, the ARIC study will be the largest population-based cohort study to evaluate the

(21)

CHAPTER II

BACKGROUND AND SIGNIFICANCE

A. Public Health Burden of Heart Failure and Obesity

Heart failure is responsible for more hospitalizations than any other condition in those 65 and older. The temporal trend of heart failure (HF) indicates a steadily increasing population burden (Masoudi, Havranek et al. 2002). Furthermore, the prevalence of heart failure is higher in U.S. blacks than in any other race group in the United States(Brown, Haldeman et al. 2005). The primary risk factor for heart failure is coronary heart disease (CHD).

Advances in medical care for CHD has resulted in longer survival time with CHD.

Therefore, this emerging heart failure epidemic may partially be the result of the enlarging population with a history of CHD, in addition to the aging of the population in

general(Braunwald 1997). Unfortunately, there is a high rate of hospital

re-admission(Krumholz, Parent et al. 1997), and a poor prognosis associated with HF, such that between 25-33 % die within the first year following an incident HF hospitalization (Croft, Giles et al. 1999; Schellenbaum, Rea et al. 2004). In 2006, the estimated cost of HF in the United States is projected to be $29.6 billion(Thom, Haase et al. 2006).

(22)

5

failure may include shortness of breath, fatigue and lower extremity edema which are not specific to heart failure. Signs of heart failure include reduced left ventricular function which can be assessed in most hospitals by echocardiography; however, the cut-point at which one defines HF is not well standardized. Originally, reduced LV function was thought to be a key component of the definition of HF, now it is understood that many people have HF with preserved LV function(Gottdiener, McClelland et al. 2002).

Risk factors for HF include advancing age, history of coronary heart disease,

hypertension, male gender, and valvular heart disease(He, Ogden et al. 2001). Overweight and obesity and its associated conditions, diabetes and even insulin resistance, have also recently been implicated as risk factors for HF(He, Ogden et al. 2001; Kenchaiah, Evans et al. 2002; Ingelsson, Sundstrom et al. 2005; Murphy, Macintyre et al. 2005). This is alarming considering the rapidly increasing prevalence of obesity(2005). Over the last few decades, the prevalence of obesity (BMI ≥ 30) in the United States has doubled. Diabetes has shown a similar increase in prevalence which is largely attributed to the obesity trends(Ford,

Williamson et al. 1997; Mokdad, Ford et al. 2000). Current estimates from the National Health and Nutrition examination survey (NHANES) show that 31 % of U.S. adults are obese and 65 % are overweight; however these estimates vary across race-gender groups. The prevalence of obesity is highest (49.6 %) in African-American women, followed by 31.3 % in white women, 28.9 % in black men and 28.7 % in white men(2005).

(23)

6

ventricular mass, a known HF predictor, than hypertension (Drazner, Rame et al. 2004). BMI is the only ponderosity metric to date to be assessed as a HF risk factor in a large population based study; the importance of measures of central adiposity (such as waist circumference and waist-hip ratio) as compared to BMI has yet to be studied in this setting. The increasing prevalence of both obesity and heart failure make this association an

important topic for further investigation.

B. Risk Factors for Heart Failure

Heart failure is a chronic disease with multiple co-morbid conditions and underlying risk factors. Many of the risk factors associated with HF are preventable. Risk factors identified in the Framingham Heart Study include advanced age, CHD, left ventricular hypertrophy, hypertension, valvular heart disease, diabetes and obesity; with hypertension having the strongest influence on the development of HF(Kannel, D'Agostino et al. 1999; Kenchaiah, Evans et al. 2002). Heart failure risk factors identified from the New Haven, Connecticut cohort for the Established Population for Epidemiologic Studies of the Elderly program were previous myocardial infarction, male gender, older age, diabetes, pulse pressure and BMI ≥

(24)

7

consistent in finding that many of the HF risk factors are also risk factor for atherosclerotic disease. It is likely that risk factor profiles differ depending on whether the etiology is ischemic or non-ischemic, although there is little data to support this. Furthermore, the distribution of risk factors for HF is believed to vary by race. Results from the National Heart Failure Project found that hypertension and diabetes were more common co-morbid conditions in blacks hospitalized with HF than whites, whereas CHD was a more common co-morbidity in whites(Rathore, Foody et al. 2003).

C. Obesity as a Risk Factor for Heart Failure

(25)

8

Figure 1. Direct and indirect mechanisms through which obesity can lead to heart failure, adapted from

RS Vasan(2003)(Vasan 2003)

Direct effects of excess adipose tissue

Independent of co-existing CHD risk factors, cardiac adaptation to excess body fat may result in decreased cardiac function(Poirier, Giles et al. 2006). This has been termed obesity cardiomyopathy. Possibly the first case study of obesity cardiomyopathy was in 1847 by William Harvey; he describes fatty adherences to the heart in an obese man with symptoms of orthopnea before his death(Ford 1950). In 1933, Smith and Willius described the

(26)

9

of pulmonary hypertension, and right heart failure(Auchincloss, Cook et al. 1955). The combination of obesity, right heart failure, and sleep apnea was termed the Pickwickian syndrome(Burwell, Robin et al. 1956). For many years, pulmonary hypertension and right heart failure (cor pulmonale) was believed the only mechanism by which obesity causes HF(Alexander 1998); however in 1965 Amad et al found increased heart weight in morbid obesity was primarily due to left ventricular hypertrophy(Amad, Brennan et al. 1965). Hemodynamic studies followed in which obesity was linked to increased plasma volume, increased cardiac output(Alexander, Dennis et al. 1962), increased LV filling pressures and decreased LV compliance (de Divitiis, Fazio et al. 1981); furthermore, improvements in these hemodynamic changes were noted after weight loss(Alexander and Peterson 1972). These findings were then confirmed in echocardiographic studies(Alpert, Terry et al. 1985; Alaud-din, Meterissian et al. 1990; Lauer, Anderson et al. 1991). Specifically, results from the Framingham Heart Study found a positive correlation between left ventricular chamber size as measured by echocardiography and severity of obesity as measured by BMI. Another study reported positive correlations of left ventricular mass with waist/hip ratio and waist circumference (Rasooly, Sasson et al. 1993).

More recently, research from animal models have lent support to the cardiotoxic effects of fat cells(Zhou, Grayburn et al. 2000). A process termed lipotoxicity has been described; it is the disruption of the usual mechanism that regulates triglyceride storage. Normally

triglycerides are stored in adipose cells, however when this process is disrupted then

(27)

10

Another novel hypothesis is that an increase in inflammatory cytokines from excess adipocytes may increase risk of HF(Vasan, Sullivan et al. 2003; McGavock, Victor et al. 2006).

Clinical manifestations of true obesity cardiomyopathy are not that common. It is thought to occur most frequently amongst those with extreme obesity (BMI ≥ 40 kg/m2) of greater than 10 years duration(Kaltman and Goldring 1976). Amongst this population, it is estimated that approximately 10 % develop circulatory congestion(Alexander, Amad et al. 1962). Therefore, the majority of heart failure associated with pre-existing obesity may be through indirect pathways.

(28)

11

Indirect effects of excess adipose tissue

More commonly, however, obesity leads to co-morbid conditions such as diabetes and hypertension that are also risk factors for heart failure. These co-morbid conditions are also known CVD risk factors; therefore, it is likely that a common pathway from obesity to heart failure is through these intermediary diseases which then cause coronary heart disease and ischemic heart failure. However, several of these secondary diseases have been found to cause heart failure independent of CHD. For example, obstructive sleep apnea causes pulmonary hypertension and resulting right heart disease which can lead to heart failure. Also, diabetes mellitus, insulin resistance and the metabolic syndrome have all been identified as independent causes of heart failure. Hypertension is known to cause left ventricular hypertrophy which can lead to heart failure.

D. Studies of Obesity as a Risk Factor for Heart Failure

Studies investigating the association of obesity with HF are reviewed in the following section. See Table 1, for a summary of these studies. Most of these studies are in primarily white populations, or specific populations (ie, clinical trial participants with specific

exclusion criteria), and BMI is the most commonly used anthropometric measure.

The Framingham Heart Study

(29)

12

committee adjudication using the Framingham Heart Study criteria for HF. There were 5,881 participants with a median follow-up of 14 years (mean age 55 years) and 496 cases of incident HF. BMI was analyzed as both continuous and categorical. After adjustment for potential confounders, the hazard ratio for incident heart failure was 1.06 (1.04-1.09 kg/m2) for a one unit change in BMI. The multi-variable adjusted HR for HF for overweight (BMI 25 – 29.9 kg/m2) and obese (BMI ≥30 kg/m2) as compared to normal weight (BMI 18.5 – 24.9 kg/m2) were 1.34 (1.08-1.67 kg/m2) and 2.04 (1.59-2.63 kg/m2), respectively. The results were similar for models with BMI and covariates treated as time-varying variables. Effect measure modification was observed for both hypertension and myocardial infarction at baseline. They found that the HR for the trend across categories of BMI was lower in those with hypertension (HR = 1.30, 1.11-1.52) than in those without hypertension, (HR = 1.66, 1.33-2.07). For those with myocardial infarction at baseline (N=148), there was no effect across categories of BMI (0.80, 0.5-1.30) for incident HF, although they were underpowered to detect a HR of 1.5 or less. For those without myocardial infarction, the HR was 1.5 (95 % CI of 1.31- 1.71). No effect measure modification was found for the

following variables: age, smoking status, gender, alcohol use, diabetes mellitus, or valvular heart disease.

The strength of this study is that it is extremely well characterized; such that the definition of heart failure is as close to a gold standard as there exists. This cohort is almost exclusively white and upper class. The size of the study is large, however may not be large enough for the inferences regarding effect modification. Effect modification was observed for

(30)

13

First National Health and Nutrition Examination Survey (NHANES I) Epidemiologic

Follow-up Study

The First National Health and Nutrition Examination Survey (NHANES I) studied the association of overweight as measured by BMI and incident heart failure(He, Ogden et al. 2001). There were 1,382 heart failure cases from 13,643 participants, age 25 to 74 years, followed for a mean of 19 years. Heart failure was defined by the presence of an ICD-9 code ‘428.0’ or ‘428.9’ from a hospitalization or death certificate. Overweight was defined as BMI ≥ 27.8 for men and ≥ 27.3 for women. The adjusted HR for overweight was 1.23 (1.00-1.52) in men and 1.34 (1.10 – 1.64) in women. These models were adjusted for age, race and time-dependent history of CHD. They determined that the population attributable fraction of HF due to overweight is 8 % (5.6 % for men and 9.6 % for women).

This strength of this study is the large size; however, it is limited by the single type of anthropometric measurement and the definition of heart failure by ICD codes. Potential effect modification by gender is an important finding.

Health, Aging and Body Composition Study

(31)

14

were present: 1) physician diagnosis of HF and treatment with diuretics and digoxin or a vasodilator, 2) pulmonary edema and cardiomegaly on chest x-ray, or 3) decreased systolic function by echocardiography or ventriculography. This study had multiple measures of body composition with dual-energy x-ray absorptiometry and computed tomography to measure visceral adipose tissue; in addition to traditional anthropometric measures of BMI, waist circumference and waist/thigh ratio were measured. All of the adiposity measures were significant predictors of heart failure in adjusted Cox models. In addition, models for waist circumference included BMI and fat mass still found waist circumference an independent predictor of HF. They also found a significant interaction for gender in terms of the waist to thigh ratio. In men the waist to thigh ratio was significantly associated with HF (HR = 1.33, 1.11-1.60), whereas in women it was not (HR = 1.06, 0.86-1.30).

Additional adjusted models considered both overweight/obese (BMI ≥ 25 kg/m2 or not) and high waist circumference (≥ 102 cm for men, ≥ 88 cm for women) as dichotomous variables. There was no significant association of overweight/obese with heart failure with BMI measured with this cut-point, however there was a significant association of high waist circumference with heart failure (HR = 1.91, 95 % CI = 1.32 – 2.75). This relationship remained significant after adjustment for obesity status. The authors concluded that it is the location of fat tissue that is important in predicting future HF; specifically, waist

circumference was the most robust predictor of HF.

(32)

15

sample size (N = 166). Although approximately 50 % of the participants were black, it is unlikely that the study was adequately powered to make inferences by race and gender. Also, this study only included those between 70-79 years of age, therefore these associations may not be generalizable to a younger age group.

The Renfrew-Paisley Study

A community based study from Renfrew and Paisley, Scotland observed 15,402 participants aged 45-64 years beginning in 1972(Murphy, Macintyre et al. 2005). Heart failure was defined by ICD codes (ICD 425.4, 425.5, 428, 402) from hospital and death records, and was not necessarily incident. BMI was the only anthropometric measure. They found 641 HF cases over 20 years of follow-up. The crude HR for a one unit increase in BMI is 1.06 (1.04-1.08). In an adjusted model including diabetes, cholesterol level, and hypertension, obesity (BMI ≥ 30 kg/m2 as compared to BMI 18.5-24.9 kg/m2) was an independent risk factor for men 2.16 (1.57-2.57) and less so for women 1.37 (1.00-1.88). This study is limited by the single type of anthropometric measurement, and the

homogeneity of this Scottish population. In addition, HF was necessarily incident. As in the NHANES study above, they found effect modification by gender.

Uppsala Longitudinal Study of Adult Men

(33)

16

(‘428’, ‘I50’ or ‘111.0’) and then validated with physician adjudication using the

classification scheme from the European Society of Cardiology(1995). Insulin sensitivity was measured with the euglycemic insulin clamp technique and anthropometric variables were BMI and waist circumference. Hazard ratios were estimated for a one standard deviation change. Multivariable models were adjusted for diabetes, prior acute MI, hypertension, smoking, left ventricular hypertrophy by electrocardiography, and serum cholesterol. They found that BMI (HR 1. 37, 95 % CI = 1.12-1.68) and waist circumference (HR 1.40, 95 % CI = 1.13-1.74) were important predictors of HF in a multivariable model, however this relationship was no longer significant when clamp glucose disposal rate was included in the model (HR for BMI = 1.17, 95 % CI = 0.92-1.50) and (HR for waist

circumference = 1.18, 95 % CI = 0.88-1.53). They found similar results sub-samples without diabetes and without obesity and in models that included interim MI.

This study was in a very specific population, elderly Swedish men, therefore it is uncertain if these results are generalizable to women, younger age groups or other ethnic groups. This is the largest study to date to report on insulin sensitivity as measured by a euglycemic clamp and heart failure. The results of this study imply that the relationship between obesity and heart failure is largely mediated by insulin resistance.

Other studies

(34)

17

criteria to the review of medical records with HF code. The HR for categorized BMI with baseline risk factors in the model was of borderline significance. This study is limited by the self-reported BMI.

Predictors of HF were studied amongst women (N =2,391) participating in the Heart and Estrogen/Progestin Replacement Study (HERS)(Bibbins-Domingo, Lin et al.). The HERS is a randomized controlled trial of hormone therapy in women with known CHD. HF was defined by hospitalization or death as discovered from routine surveillance and then validated by committee adjudication. There were 237 cases over a mean follow-up of 6.3 years. The adjusted HR for obese participants (BMI >35 kg/m2) compared to those of normal weight (BMI 18.5 - 25 kg/m2) was 1.9 (95 % CI = 1.1-3.0). This study is small and includes only women with known CHD. Also, this is a population participating in a randomized controlled trial which is likely healthier than the general population.

The Heart Outcomes Prevention Evaluation (HOPE) study is a randomized controlled trial assessing the use of an angiotensin-converting enzyme inhibitor (ramipril), vitamin E, or their combination in reducing CVD events(Dagenais, Yi et al. 2005). This study only

(35)

18

added to the model these findings were no longer significant. There are two main limitations of this study which reduce it’s generalizability: 1) the use of tertiles to define cut-points for the main exposures are not recommended especially when recommended category

boundaries are available(Greenland and Rothman 1998), and 2) the population is high-risk male participants in an randomized controlled trial (RCT).

Summary

Nearly all of these studies find a significant association between obesity and heart failure. Only three studies have waist circumference in addition to BMI; however, these particular studies are not population-based(Dagenais, Yi et al. 2005; Ingelsson, Sundstrom et al. 2005; Nicklas, Cesari et al. 2006), and instead are in very specifically defined populations.

(36)

19

Table 1. Summary table of studies of obesity as a predictor of heart failure

Author, year Study Population Anthropometric

Measures

Outcome Definition Limitations

Kenchaiah, 2002 491 cases from 5,881 participants, cohort study from

Framingham, Mass., mostly white, mean age 55

BMI Adjudicated incident heart

failure

BMI only metric, largely white

cohort

Chen, 1999 173 cases from 1,749 participants from the New Haven,

Connecticut cohort in the Established Population for Epidemiologic Studies of the Elderly, community based

cohort, aged 65 years and older, free of CHD, 18 % black participants

BMI, based on self report of height and

weight

Incident hospitalized heart failure based on chart

review

Self reported BMI, mostly white cohort

Nicklas, 2006 166 cases from 2,435 participants in the Health ABC

(Aging and Body Composition) study, cohort study of those 70-79 years without CHD, biracial (black/white)

BMI, waist circumference, waist/thigh ratio, dual-energy x-ray absorptiometry, CT scans

Adjudicated incident heart failure

Limited sample size, free of CHD

at baseline

He, 2001 1,382 cases from13,643 participants in First National

Health and Nutrition Examination Survey (NHANES), 25 to 74 years of age

BMI First listing of ICD-9 code

‘428’ from a hospitalization or UCOD on death

certificate

BMI only metric, non-adjudicated definition for HF

Murphy, 2005 641 cases from 15,402 participants in the

Renfrew-Paisley study, 45-64 years of age from Scotland

BMI Heart failure ICD codes

from a hospitalization or death certificate

BMI only metric, non-adjudicated definition for HF

Bibbins-Domingo, 2004

237 cases from 2,391 women with known CHD participating in the Heart and Estrogen Replacement Study (HERS), randomized controlled trial of estrogen

replacement therapy

BMI Adjudicated incident heart

failure

BMI only metric, only women,

RCT, small sample

Dagenais, 2005 297 cases from 8,802 participants in the Heart Outcomes

Prevention Evaluation (HOPE) study, high risk patients for CVD events, 75 % male, randomized controlled trial

of ACE inhibitor (ramipril) and vitamin E

BMI, waist circumference,

waist/hip ratio

Adjudicated incident heart failure

RCT, included only high risk

patients

Ingelsson, 2005 104 cases from 1,187 participants in the Uppsala

longitudinal study of adult men, a community based observational cohort from Sweden, 70 years and older

BMI, waist circumference

First hospitalized heart failure, adjudicated

(37)

20

E. Validation Studies of ICD codes to Define Heart Failure

Hospital discharge codes are based on the International Classification of Diseases (ICD) system which was created primarily for administrative rather than epidemiologic purposes. Despite this, ICD codes are often used in large epidemiologic studies to define disease, when the gold standard may not be feasible(He, Ogden et al. 2001; Murphy, Macintyre et al. 2005). Unfortunately, the gold standard of committee adjudication is expensive and time

consuming; also, there are not clearly agreed upon criteria to define heart failure(Goldberg and Konstam 1999). One reason for misclassification from the use of ICD codes are their linkage to hospital reimbursement. In particular, heart failure as a complication during a hospitalization can significantly increase the Medicare reimbursement associated with that hospitalization(Psaty, Boineau et al. 1999). Therefore, there is a financial incentive for hospitals to “up-code” or to list a heart failure ICD codes. As a result, studies have tried to determine the amount of misclassification of heart failure when diagnosed by ICD codes. These validation studies are reviewed below (see summary in Table 2).

Corpus Christi Heart Project

The Corpus Christi Heart Project is a population-based study with the main goal of

surveillance for hospitalized coronary heart disease; therefore, these hospitalizations included those for definite and possible myocardial infarction, aortocoronary bypass surgery, and transluminal coronary angioplasty(Goff, Pandey et al. 2000). This data was used to

(38)

21

criteria of a physician diagnosis of acute heart failure or evidence of pulmonary edema on chest x-ray.

The overall prevalence of a heart failure ICD code was (1197/5083) 23.5 %, whereas the most frequent heart failure ICD code was ‘428’ with a prevalence of 20.4 % (1,035/5,083), this was followed by 402.x1 (hypertensive heart disease with congestive heart failure) with a prevalence of 2.6 %. All other heart failure ICD codes had a prevalence of less than 1 %. The test characteristics of code ‘428’ were as follows: sensitivity = 62.8 % (864/1376), specificity = 95.4 % (3536/3707), positive predictive value = 83.5 % and negative predictive value = 87.4 %; whereas, the test characteristics for any HF ICD code were: sensitivity = 67.1 % (923/1376), specificity = 92.6 % (3433/3707), positive predictive value = 77.1 %, and a negative predictive value = 88.3%. The sensitivity was higher, but the specificity was lower when all of heart failure ICD codes were used.

This study is a large population-based validation study of ICD codes for heart failure in those with a co-existing ICD code for coronary heart disease; therefore, it is likely that the resulting test characteristics are not generalizable those without CHD. It is notable that ICD code ‘428’ is the most frequently used HF code and it is more specific than when combined with all other HF codes.

The Cardiovascular Health Study

(39)

22

398.91 (rheumatic heart failure), 425 (cardiomyopathy), 402.01, 402.11, 402.91

(hypertension with heart failure), and 997.1 (cardiac failure postoperatively). They only included those that survived hospitalization. Of these discharge codes, code ‘428’ was the most frequent at 70.6 percent of the events followed by code 997.1 for “cardiac failure postoperatively” (11.6 %). These potential cases were then validated with committee adjudication using criteria created by the CHS. The CHS criteria required a physician diagnosis of heart failure in addition to supporting evidence of any of the following: heart failure signs and symptoms, pulmonary edema on chest x-ray or evidence of treatment for heart failure.

Using only the ICD-9 code ‘428’, there were 523 cases that were adjudicated as heart failure, leaving 364 individuals with a code ‘428’ that were not considered heart failure after adjudication. The test characteristics of a code ‘428’ by the CHS criteria are as follows: sensitivity = 0.71, specificity = 0.925, positive predictive value = 0.59 and negative

predictive value = 0.96. Mortality was selected as a surrogate endpoint for predictive validity of these HF identification methods. Mortality was not significantly different across these two methods of HF event definition. Also notable considering the proposed dissertation, BMI did not vary across the type of event definition (ie adjudication vs. discharge diagnosis).

This is a well done validation study from a large population-based cohort. The only issues that decrease generalizability to other population-based studies are the older age group (65 years and older) and the exclusion of those who died during hospitalization. It is

(40)

23

Olmsted County, Minnesota

A study from Olmsted County, Minnesota primarily on incidence and survival also tested the accuracy of ICD codes for a heart failure by validation with 2 sets of criteria,

Framingham and specific clinical criteria.(Roger, Weston et al. 2004) This is community-based open cohort with 22 years of follow-up. There were a large proportion of potential cases (26 %) that were outpatient only. They found 7,298 (80%) of those identified as potential cases had an ICD code ‘428’, whereas only 1,877 potential cases were identified by other HF codes without an accompanying ‘428’ (these other codes were 402.01,

‘hypertensive heart disease malignant with congestive heart failure’, 402.11, ‘hypertensive heart disease benign with congestive heart failure’, 425, ‘cardiomyopathy’, 429.3,

‘cardiomegaly’, and 514, ‘pulmonary congestion’). First, they validated these potential cases by applying the Framingham criteria using committee adjudication. They found that 82 % of potential cases with a ‘428’ code met Framingham criteria, whereas only 14-30 % of the other HF ICD codes (without an associated ‘428’) met Framingham HF criteria. Next, they validated with clinical criteria of a physician diagnosis of heart failure. In this case, 90% of code ‘428’ had a physician diagnosis of heart failure, whereas only 14-36 % of the other codes (without a ‘428’) met clinical criteria. They found that secular trends for incidence and survival were similar between the Framingham and clinical criteria.

(41)

24

Women’s Health Initiative (WHI)

A study from the observational cohort of the Women’s Health Initiative (WHI) tested the accuracy of diagnostic codes to local adjudication(Heckbert, Kooperberg et al.). Heart failure was defined by the following criteria: symptoms and signs of HF and either pulmonary edema by chest x-ray, documented dilated ventricle or decreased ventricular function, or physician diagnosis of HF and receiving medical treatment(Curb, McTiernan et al. 2003). There were 93,676 post-menopausal women in the observational cohort of which 1,241 were hospitalized with an ICD 9 code ‘428’; of these codes, only 603 were validated as HF by local adjudication using a standard definition developed by the WHI. The positive predictive value of an ICD code 428 was estimated as 48.6 %. They also validated ICD code 425 (N = 134) and found the PPV was a little lower at 45 %. The main limitation of this study was that participants were all female; in addition, participants were mostly upper class and from urban areas. Compared to other validation studies, this one was not population-based.

National Registry for Atrial Fibrillation

A study from the National Registry for Atrial Fibrillation II evaluated the accuracy of ICD codes for co-morbid conditions including HF(Birman-Deych, Waterman et al. 2005). This registry includes patient information from 23,657 Medicare beneficiaries representing 3,586 hospitals in all 50 states. Mean age was 79 years of age. There were 11,014 heart failure cases based on medical record review for both chronic and current heart failure. They

(42)

25

estimated the following test characteristics based on current or past HF events: sensitivity = 0.76, specificity = 0.97, positive predictive value = 0.97 and negative predictive value = 0.74. When past HF was excluded, then the specificity (0.86) and PPV (0.85) were lower, and the sensitivity was higher (0.83).

One would expect a higher positive predictive value for a heart failure ICD code in a population with atrial fibrillation, since this arrhythmia is even a component of some HF diagnostic criteria(Eriksson, Caidhal et al. 1987). Unfortunately, test characteristics were not presented for ICD code ‘428’ only.

Summary

These five validation studies provide insight into misclassification rates of ICD codes for heart failure. Unfortunately, there is no gold standard to define heart failure; therefore, the diagnostic criteria vary by study. In the proposed dissertation, there are no internal validation studies from which to develop these estimates. Four of the above studies found that ICD code ‘428’ was by far the most frequently documented ICD code for heart failure(Goff, Pandey et al. 2000; Heckbert, Kooperberg et al. 2004; Roger, Weston et al. 2004;

(43)

26

(44)

27

Table 2. Summary table of validation studies of ICD codes to define heart failure

First author, year Study Population Validation criteria Results

Goff, 2000 Corpus Christi Heart Project, included

hospitalizations for definite or possible MI, aortocoronary bypass surgery, and transluminal coronary angioplasty, population-based, 5,083 hospitalizations with possible heart failure

Physician diagnosis of acute HF or pulmonary edema on chest x-ray

For ICD code ‘428’:

Sensitivity: 62.8 % (864/1376) Specificity: 95.4 % (3536/3707) PPV: 83.5 % (1035)

NPV: 87.4 % (4048)

Schellenbaum, 2005 Cardiovascular Health Study, population-based

study, 65 years and older, 1,209 with possible incident heart failure

Physician diagnosis of HF, in addition to any of the following: HF signs or symptoms, pulmonary edema on chest x-ray or evidence of treatment for HF

For ICD code ‘428’: Sensitivity: 71 % Specificity: 93 % PPV: 59 % NPV: 96 %

Roger, 2004 Olmsted county, Minnesota, population-based

open cohort, included 26 % outpatient only cases, 4,537 possible incident heart failure cases

Tested 2 criteria Framingham and

Physician diagnosis of HF

For ICD code ‘428’: Framingham criteria: PPV: 82%

Clinical criteria: PPV: 90%

Heckbert, 2004 Women’s Health Initiative, 93,657

post-menopausal women in the observational cohort, 1,241 participants with a code ‘428’

Symptoms/signs of HF and either pulmonary edema by chest x-ray, documented dilated ventricle/decreased ventricular function, or physician diagnosis of HF and receiving medical treatment

For ICD code ‘428’: PPV: 49 %

For all ICD codes: Sensitivity: 79 % PPV: 45 %

Birman-Deych, 2005 National Registry for Atrial Fibrillation II,

23,657 Medicare beneficiaries with an ICD code for atrial fibrillation, incident and recurrent HF

Chart review for mention of current or history of heart failure

(45)

28

F. Methods to Adjust for Bias

Some degree of misclassification is common in observational studies; however, bias due to misclassification is rarely quantified. The conventional approach to the presentation of uncertainty in the scientific literature is the 95 % confidence interval; although, confidence intervals only quantify residual random error(Poole 2001). Unfortunately, the confidence interval is often misinterpreted to represent all sources of error (Greenland 2001). Currently, the standards for publishing in most journals do not include the need to go beyond the

uncertainty of random error. Sensitivity analyses attempt to quantify the effect of bias on the results of a study. These techniques have not yet reached the mainstream; however, several epidemiologists have called for a more thorough presentation of the uncertainty inherent in scientific research(Maclure and Schneeweiss 2001; Greenland 2005). In fact, a journal named ‘Epidemiologic Perspectives & Innovations’ was created in 2004 inviting the

exploration of such issues(Phillips, Goodman et al. 2004). The editors request submission of research that expresses “full and proper disclosure of uncertainty in study results” and

explores “decision making in the face of this full disclosure”(Maldonado and Phillips 2004). A simple sensitivity analysis for misclassification is to back-calculate the expected results given a plausible set of specific estimates for sensitivity and specificity(Greenland 1998). This method does not consider the likelihood of each set of values(Phillips 2003). In addition, the presentation of such results requires a table rather than a more succinct

(46)

29

type of incorporation of uncertainty into regression analysis is called a Monte Carlo Risk Assessment (MCRA) or Monte Carlo uncertainty analysis. Such an analysis requires the statement of a prior distribution for each sensitivity parameter which should include an explicit rationale for the priors(Greenland 2001). More specifically, an MCRA can

iteratively sample for multiple pairs of sensitivity parameters from the input distribution, then create modified datasets for each set of parameters, and then re-run the conventional

regression for each modified dataset. A summary of the distribution of the effect estimates obtained from each modified dataset can be presented graphically, or with 95 % “uncertainty intervals” by taking the 2.5th and 97.5th percentiles from this distribution.

This is an improvement over older sensitivity analysis methods for the following reasons: 1) a distribution of sensitivities and specificities can be specified, such that the likelihood of each set of values is taken into account; 2) the distribution of sensitivities and specificities are used to create modified datasets for analysis with conventional regression, such that a summary distribution of effect estimates is created; 3) concise summarization of results can be presented that account for random error, systematic error and both; 3) the magnitude of multiple sources of bias (i.e., from the exposure, outcome or covariates) can be considered in one analysis that still results in concise summary statistics and graphs. This technique is useful for the proposed study as we would like to consider the effect of misclassification of the outcome (incident heart failure). After specifying a distribution for the bias due to outcome misclassification, then we will have a new summary statistics to compare with the conventional analysis.

In an article by Lash and Fink(Lash and Fink 2003) entitled, ‘semi-automatic

(47)

30

multiple biases. They evaluated 3 types of systematic error (selection bias, misclassification, and confounding). These three types of systematic error are analyzed independently and all together; next random error is incorporated into each analysis (See below). The end result is the inclusion of all sources of error in the bottom right three columns such that the simulated median is 1.52 with 95 % uncertainty intervals of 0.81-2.81; which can be compared to the conventional hazard ratio of 2.0 with 95 % CI of 1.2-3.4. The relative impact of each source of error can be visualized in the below in Error! Reference source not found.. In this figure, the dotted/dashed line represents the results from the sensitivity analysis including all 3 sources of systematic error, the dotted line represents the results of the conventional analysis, and the solid line represents the combination of systematic and random error from the bootstrapped samples. The availability of high performance software has facilitated this advancement in multiple bias modeling, however such techniques are not often used. We suggest the use of this methodology for the analysis of systematic error in the proposed study. From Error! Reference source not found. and

Table 3, one can envision how the results of such an analysis might appear.

(48)

31

Figure 3. Graphic representation of the results summarized in

(49)
(50)

33

G. Impact measures of risk factor-disease associations

In epidemiology we most often study the association of a given exposure with a disease. Once a potentially causative association is found with a modifiable exposure then we want to determine the burden of disease that could be prevented given the implementation of

available interventions to modify the exposure. This latter part allows prediction of the possible public health impact of programs to reduce or eliminate a given exposure. The primary impact measure used in practice is the population attributable fraction (also known as, attributable risk). The attributable fraction, introduced by Levin in 1953(Levin 1953), is often misused and/or misinterpreted(Rockhill, Newman et al. 1998). It estimates the

proportional reduction in disease given complete elimination of an exposure. For many exposures complete elimination is impossible or highly unlikely(Benichou 2007). Obesity and overweight is an exposure that is unlikely to be eliminated. Despite this, the attributable fraction has been used extensively in the recent articles attempting to estimate the burden of deaths attributable to overweight and obesity(Flegal, Graubard et al. 2004; Flegal, Graubard et al. 2005). A more useful calculation for risk factors such as obesity is the generalized impact fraction, also known as the potential impact fraction and the generalized attributable fraction(Benichou 2001; Rodgers, Ezzati et al. 2004). It is a generalization of the attributable fraction which estimates the proportional reduction in disease incidence given a reduction in prevalence of a given risk factor. These calculations are often stratified by important

(51)

34

impact of intervention will likely reveal important changes in the burden of disease even when risk factor-disease associations are relatively weak(Morgenstern and Bursic 1982).

Generalized impact fraction

The generalized impact fraction was originally described by Walter in 1980(Walter 1980), then further elucidated and coined by Morgenstern and Bursic in 1982(Morgenstern and Bursic 1982). It is “the proportional reduction in the total number of new (incident) cases of a certain disease, resulting from a specific change in the distribution of a risk factor in the population at risk.”(Morgenstern and Bursic 1982). Despite its introduction over 25 years ago, it hasn’t caught on as either a replacement for or as an additional measure to report with the attributable fraction.

Morgenstern and Bursic (1982) illustrate the use of the generalized impact fraction with an example of the impact of a hypothetical weight loss programs on the incidence of

(52)

35

are not secular trends in risk of disease that are not due to the intervention; and there are not secular trends in intermediary diseases, such as diabetes for obesity, that would effect the relationship between the exposure and disease.

Inherent in estimating the GIF, one must consider a counterfactual approach(Rodgers 2002). The theoretical minimum risk distribution would be the complete elimination of obesity, or for a continuous measure it would be the point of lowest risk on the distribution. However, it is more likely that there could be partial eradication of obesity. There are

multiple counterfactual situations that could be considered between the current and minimum distributions and these are called distributional transitions by the World Health Organization. Most likely to occur are small distributional shifts, such as 10 % or 20 % change. Murray and Lopez introduce four types of counterfactual exposure distributions: theoretical

minimum risk, plausible minimum risk, cost-effective minimum risk, and feasible minimum risk. In our case, we are interested in feasible minimum risk(Murray and Lopez 1999). Feasible means that the distribution change has been achieved in some population and is also possible for the current population. A plausible distribution is imaginable, but rather may be possible for some society during some time period.

Feasible goals for weight reduction

(53)

36

develop US national, state and community plans for health improvement. Specifically, their objective is to reduce the proportion of obese adults aged 20 years and older to 15 % from 23% (25 % in females and 20 % in males, data from 1988-1994, age adjusted to year 2000). A second objective is to increase the proportion of adults at a healthy weight to 60 % from 42 % (45 % in women, 38 % in men, data from 1988-1994, age-adjusted to year 2000).

Obesity and overweight are unfortunately increasing rather than decreasing(Rosamond, Flegal et al. 2007). Many strategies for weight reduction exist, however all have limited effectiveness. A meta-analysis by Franz et al categorized weight loss clinical trials into 8 types of intervention: diet and exercise, exercise alone, advice alone, meal replacements, very-low-energy diets, and weight-loss medications (orlistat and sibutramine)(Franz, VanWormer et al. 2007). Results from this meta-analysis, which includes clinical trial completers with at least one year of follow-up, found that the mean weight loss was 5-8.5 kg (5-9%) at 6 months and stabilized at 4.5-7.5 kg (4.8-8 %) at 12 months. A diet and exercise intervention would be most applicable for a population wide strategy. Those in the diet and exercise group lost 7.6 kg at 12 months, whereas those in the diet-alone group lost 4.6 kg. A systematic review of diet and exercise trials reported that a mean of 6.7 kg of weight loss was maintained after one year(Curioni and Lourenco 2005). Based on these findings of those who completed one year of follow-up, it is apparent that at best only a modest weight reduction could be expected from a population wide intervention.

H. Summary and public health significance

(54)

37

of these studies used BMI as the sole metric of adiposity and were in primarily white populations, specific clinical trial populations or isolated community populations. Insulin resistance is likely an important intermediary factor in the association between obesity and HF. Waist hip ratio, a metric of central adiposity, is more closely associated with insulin resistance than BMI. Selecting the best anthropometric for the prediction of HF could have public health implications for the screening and prevention of HF. Currently, screening for obesity and overweight is primarily assessed with BMI, rather than waist circumference or WHR. Further evidence supporting measures of central adiposity, rather than BMI, could eventually result in a shift in current practice patterns.

The magnitude of misclassification of heart failure due to the use of ICD codes has been quantified in several studies. For large observational studies without committee validation, one method to adjust for disease misclassification is to perform a sensitivity analysis to show the effect of misclassification. Approaches called multiple bias modeling allow the

incorporation of various types of bias, either from the main exposure, outcome or covariates, into multivariate modeling using Monte Carlo techniques. This study provides further example of how Monte Carlo Risk Assessment techniques can be used to succinctly

summarize the possible effects of systematic error. The application of this method has public health importance in providing further example of how the interpretation of results can easily include a sensitivity analysis which incorporates systematic error in addition to random error. For studies in which results will potentially change policy or alter patient care, it is vital that such estimates of uncertainty be considered in the interpretation of results.

We estimated the public health impact on the incidence of heart failure given a

(55)

38

(56)

CHAPTER III

METHODS

A. Study population

The ARIC cohort was recruited using probability sampling of those aged 45-64 from the following four US communities: Forsyth County, North Carolina (includes Winston-Salem); the city of Jackson, Mississippi; the northwestern suburbs of Minneapolis, Minnesota; and Washington County, Maryland (includes Hagerstown). The distribution of blacks and whites from each county is representative of the area (mostly white in Minneapolis and Washington County), except for Forsyth County in which blacks were over-sampled (15 %) and in

Jackson where only blacks were sampled. Response rates were 46 % in Jackson and between 65-67 % for the other communities. The design and rationale of the ARIC study

(Investigators 1989) and the comparison between responders and non-responders (Jackson, Chambless et al. 1996) have been previously published. The institutional review boards from each site approved the ARIC study; also, the institutional review board at UNC-Chapel Hill approved this dissertation. All participants provided written informed consent.

(57)

40

participants to the local field center. Staff members were certified in the appropriate method for obtaining consent. The first portion of the examination was performed following a 12 hour fast. Cohort examinations took place every 3 years for 4 visits, beginning with the baseline visit in 1987-1989. Between cohort examinations, a telephone questionnaire was administered yearly to identify intervening hospitalizations and deaths. In addition, community-wide surveillance was performed to identify all cohort hospitalizations and deaths(Investigators 1989).

B. Exclusion criteria

Racial groups not classified as white or black (N=48) and those missing anthropometry (N=33) were excluded. In addition, those with prevalent heart failure at baseline were excluded from this analysis by the following criteria: 1) those answering “yes” to the following question: “Were any of the medications you took during the last two weeks for heart failure?” (N = 83), or else 2) those with stage 3 or ‘manifest heart failure’ by applying Gothenburg criteria (N= 699), or 3) those who did not meet one of these 2 criteria, but were unresponsive to the HF medication question or any component of the Gothenburg criteria (N = 289). The Gothenburg criteria are based on a study from Gothenburg, Sweden of men born in 1913(Eriksson, Caidahl et al. 1987). It is composed of three scores: 1) cardiac, 2)

pulmonary, and 3) therapy. In order to have stage 3 heart failure, one must have a point from each category. See Table 4, for a description of the Gothenburg criteria. All current

(58)

41

standardized methodology(Vitelli, Crow et al. 1998). All other components were determined by participant self-report. After these exclusions, the total sample size was 14,690.

Table 4. Description of Gothenburg score Gothenburg components

Cardiac Coronary Heart Disease -1 point if ever, 2 points if within the last year Angina - 1 point if ever, 2 points if within the last year

Leg edema – 1 point

Shortness of breath at night – 1 point Rales on lung exam – 1 point

Atrial fibrillation on ECG – 1 point Pulmonary History of bronchitis – 1 point

History of asthma – 1 point

Cough, phlegm or wheezing – 1 point Rhonchi on lung exam – 1 point Therapy Treatment with digoxin – 1 point

Treatment with diuretics – 1 point

C. Ascertainment of heart failure events

(59)

42

HF. Therefore, many deaths due to HF would be missed using only a definition of HF as underlying cause of death.

D. Incident heart failure event criteria

Incident HF was defined as the first occurrence of either: 1) a hospitalization which included an ICD-9-CM (International Classification of Diseases, 9th revision, clinical modification) discharge code beginning with ‘428’ in any position (N = 1,329) or else 2) a death certificate with an ICD-9 code beginning with ‘428’ (HF) or ICD-10 code ‘I50’ (HF) in any position (N = 76). Follow-up time for those with an incident HF event was defined as the time from the date of their baseline examination (1987-1989) until the incident event (follow-up through Dec. 31st, 2003). The end of follow-up time for those without HF was the first occurrence of either: 1) December 31st, 2003, 2) date of last contact for those lost to follow-up, or 3) date of death.

E. Anthropometric measures

(60)

43

muscles. Values were rounded down and were recorded to the centimeter. Weight

measurements were performed using a scale (Detecto model 437) that was zeroed daily and calibrated quarterly. Body mass index (BMI) was calculated as weight divided by height squared (kilograms/meters2), whereas waist/hip ratio is the waist girth divided by the hip girth. Inter-technician reliability coefficients for waist and hip girth and WHR were > 0.91.(Ferrario, Carpenter et al. 1995)

F. Baseline covariate definitions

(61)

44

diabetes, medication use for diabetes over the last 2 weeks, a blood glucose ≥ 126 mg/dl fasting or a blood glucose ≥ 200 mg/dl non-fasting. Forced expiratory volume (FEV1) was obtained from pulmonary spirometry performed by trained technicians with computer assistance. The FEV1 measurements have been adjusted for age, race, gender and height(Shahar, Boland et al.). Methods for the measurement of blood levels of albumin, creatinine, and glucose have been previously described (Eckfeldt, Chambless et al. 1994). Low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), triglycerides, and total cholesterol were measured in fasting state using standardized methods (Investigators 1989).

G. Data quality

Data was directly entered into a computer-assisted data collection system. Suspicious values were immediately detected while the participant was onsite, such that the value could be confirmed or corrected. The main study data was held at the coordinating center; the main data was updated weekly by diskettes of study data mailed from the centers. Reports on data were generated routinely for review by all study sites. Quality control measures were

(62)

45

included: 1) assessment of participant effort during spirometry; 2) assessment for digit preference in blood pressure data; 3) measurement of blood flow rate during venipuncture; 4) repeat anthropometric measures by the same and different technicians; and 5) blind analysis of duplicate blood samples and electrocardiograms(Investigators 1989).

H. Statistical power analysis

Since there are accepted cut-points for BMI, we used BMI as the main exposure for the statistical power analysis. The data for this study has already been collected and therefore the numbers per groups has already been defined. Power analyses for incident HF were estimated stratified by gender for the comparison between the highest category of BMI (BMI >30) as compared to the referent group (BMI < 25). Power analysis assumes no

confounding, no missing values, and no bias. Preliminary data from the ARIC study was used to determine the exponential survival parameters for each BMI group, whereas we assumed a dropout rate of 5 % per year. The group with the smallest number was provided as “n per group” and was held constant across the table. The highlighted numbers were estimated by the software.

(63)

46

power Error! Reference source not found.) based on a HR of 2.9 is 99 % if there were 66 cases (example 1), whereas as when the HR decreases to 1.5, the power decreases to 91 % if there were 275 cases (example 3). The actual total number of events for these two BMI categories combined is 372 for men and 383 for women. Since we have over 350 cases for both men and women, this tables shows that we have adequate power even after stratification by gender. Power estimates and tables were created using NQuery Advisor 5.0.

Table 5. Two group test of equal exponential survival, with exponential dropout, only men included

1 2 3 4

Test significance level, α 0.050 0.050 0.050 0.050

1 or 2 sided test? 2 2 2 2

Length of accrual period 3.00 3.00 3.00 3.00

Maximum length of follow-up 16.00 16.00 16.00 16.00

Common exponential dropout rate, d 0.0500 0.0500 0.0500 0.0500

BMI >30, exponential parameter, λ 1 0.0101 0.0101 0.0101 0.0101 BMI <25, exponential parameter, λ 2 0.0048 0.0067 0.0078 0.0072

Hazard ratio, h= λ 1 / λ 2 2.104 1.500 1.300 1.400

Power ( % ) 99 89 57 77

n per group 1567 1567 1567 1567

Total number of events required, E 133 248 268 258

Table 6. Two group test of equal exponential survival, with exponential dropout, women only

1 2 3 4

Test significance level, α 0.050 0.050 0.050 0.050

1 or 2 sided test? 2 2 2 2

Length of accrual period 3.00 3.00 3.00 3.00

Maximum length of follow-up 16.00 16.00 16.00 16.00

Common exponential dropout rate, d 0.0500 0.0500 0.0500 0.0500

BMI >30, exponential parameter, λ1 0.0072 0.0072 0.0072 0.0072 BMI <25, exponential parameter, λ 2 0.0025 0.0036 0.0048 0.0055 Hazard ratio, h= λ 1 / λ 2 2.880 2.000 1.500 1.300

Power ( % ) 99 99 91 62

n per group 2405 2405 2405 2405

(64)

47

(65)

48

Table 7. Estimate of power* to assess for multiplicative effect measure modification by the following variables, given probabilities and sample size in the ARIC

Power in Men Power in Women

Age 0.31 0.34

Race 0.49 0.35

Hypertension 0.40 0.32

Diabetes 0.25 0.16

CHD 0.19 0.07

I. Statistical Analysis

The distributions of all three exposures (BMI, WHR, and WC) were inspected for outliers. The distribution of men and women were compared for each exposure. Pearson correlation coefficients were estimated to determine the correlation between BMI, waist-hip ratio and waist circumference. Categorization of BMI was defined as established in the literature(NIH 1998): 1) BMI < 25 kg/m2 (normal weight); 2) BMI between 25-30 kg/m2 (overweight); and 3) BMI >30 kg/m2 (obese). No well accepted epidemiologic standard exists for the

categorization of WHR and WC, therefore both of these measures were categorized into approximate tertiles by gender. Although a sex-specific dichotomous cut-point has been suggested for WC (obese defined as men > 102 cm and women > 88 cm)(NIH 1998), we preferred approximate sex-specific tertiles of WC for better specification. Categorized variables were represented as indicator variables with comparison to the lowest group (normal weight) as referent. We evaluated BMI in classes of weight as represented in the clinical guidelines from the National Institutes of Health(1998). In addition, we

Figure

Figure 1.  Direct and indirect mechanisms through which obesity can lead to heart failure, adapted from  RS Vasan(2003)(Vasan 2003)
Figure 2.  Depiction of lipotoxicity, adapted from McGavock JM (2006)(McGavock, Victor et al
Table 1.  Summary table of studies of obesity as a predictor of heart failure
Table 2. Summary table of validation studies of ICD codes to define heart failure
+7

References

Related documents

From the case studies, four social and organizational characteristics that impact the functioning of multidisciplinary building design teams are discussed –

At the very least, the Health Department should have sent someone out to the hospital to assess Patient X and create a plan for after care – standard protocol, according to

[78], presented in this literature a control strategy is proposed to regulate the voltage across the FCs at their respective reference voltage levels by swapping the switching

This is important because the evidence from the OECD’s Programme for International Student Assessment (PISA) is not only that Australian literacy and numeracy levels at 15 years of

Survival analysis according to the treatment for the recurrent disease showed that local therapy (repeat hepatectomy or ablation) for intrahepatic recurrence improved OS compared

Furthermore, we found that the conventional RF spoiling scheme that is typically used for SPGR and STFR does not always work when pulses with crushers like fat sat are applied, so

The SAVE U trial is designed to compare sacrospinous fixation with vaginal hysterectomy in the treatment of uterine prolapse stage 2 or higher in terms of prolapse recurrence,